Reply on RC1

Опубликована: Июнь 20, 2024

Abstract. In this study, we used deep learning models with recurrent structure neural networks to simulate large-scale groundwater level (GWL) fluctuations in northern France. We developed a multi-station collective training for GWL simulations, using both “dynamic” variables (i.e. climatic) and static aquifer characteristics. This approach offers the possibility of incorporating dynamic features cover more reservoir heterogeneities study area. Further, investigated performance relevant feature extraction techniques such as clustering wavelet transform decomposition, intending simplify network regionalised information. Several modelling tests were conducted. Models specifically trained on different types GWL, clustered based spectral properties data, performed significantly better than whole dataset. Clustering-based reduces complexity data targets information efficiently. Applying without prior can lead learn dominant station behavior preferentially, ignoring unique local variations. respect, pre-processing was found partially compensate clustering, bringing out common temporal characteristics shared by all available time series even when these are “hidden” because too small amplitude. When employed along thanks its capability capturing essential across scales (high low), decomposition technique provided significant improvement model performance, particularly GWLs dominated low-frequency advances our understanding simulation learning, highlighting importance approaches, potential preprocessing, value attributes.

Язык: Английский

---

Spatial and temporal forecasting of groundwater anomalies in complex aquifer undergoing climate and land use change

Journal of Hydrology, Год журнала: 2024, Номер 639, С. 131525 - 131525

Опубликована: Июнь 21, 2024

Monitoring groundwater (GW) level variations, or anomalies in multiple wells, over long periods of time is essential to understanding changes regional resource availability. However, it challenging predict these GW the term agricultural areas due complicated boundary conditions, heterogeneous hydrogeological characteristics, and extraction, as well nonlinear interactions among factors. To overcome this challenge, we developed an advanced modeling framework based on a recurrent neural network short-term memory (LSTM) alternative complex computationally expensive physical models. were forecast two months advance (t + 2) evaluation drivers that influence dynamics densely irrigated regions. An application new approach was conducted Wisconsin Central Sands (WCS) region U.S., one most productive The for period 1958–2020 by utilizing easily accessible dynamic static variables represent hydrometeorological geological characteristics. anomaly observations acquired from 26 piezometers (wells) installed sandy aquifer WCS 10–60 years. subset ∼ years, not used model training, can out with coefficient determination R2 0.8. Additionally, MAE less than 0.34 m/month across study both temporal spatial frameworks. Groundwater showed high spatiotemporal variability, their responses are influenced differently catchment geology, climate, topography locations. Sites higher autocorrelation previous two-months reduced bias increasing R2. Land use change irrigation pumping have interactive effects forecasting. novelty identifying fluxes. This case-specific information location-related simplification, modification, assumption LSTM unique contribution existing literature. Our be method simulating water availability where subsurface properties unknown difficult determine.

Язык: Английский

---

Reply on RC2

Опубликована: Июль 25, 2024

Abstract. In this study, we used deep learning models with recurrent structure neural networks to simulate large-scale groundwater level (GWL) fluctuations in northern France. We developed a multi-station collective training for GWL simulations, using both “dynamic” variables (i.e. climatic) and static aquifer characteristics. This approach offers the possibility of incorporating dynamic features cover more reservoir heterogeneities study area. Further, investigated performance relevant feature extraction techniques such as clustering wavelet transform decomposition, intending simplify network regionalised information. Several modelling tests were conducted. Models specifically trained on different types GWL, clustered based spectral properties data, performed significantly better than whole dataset. Clustering-based reduces complexity data targets information efficiently. Applying without prior can lead learn dominant station behavior preferentially, ignoring unique local variations. respect, pre-processing was found partially compensate clustering, bringing out common temporal characteristics shared by all available time series even when these are “hidden” because too small amplitude. When employed along thanks its capability capturing essential across scales (high low), decomposition technique provided significant improvement model performance, particularly GWLs dominated low-frequency advances our understanding simulation learning, highlighting importance approaches, potential preprocessing, value attributes.

Язык: Английский

---

Groundwater level projections for aquifers affected by annual to decadal hydroclimate variations

Authorea (Authorea), Год журнала: 2024, Номер unknown

Опубликована: Сен. 2, 2024

In a context where anticipating future trends and long-term variations in water resources is crucial, improving our knowledge about most types of aquifer responses to climate variability change necessary. Aquifers with dominated by seasonal (marked annual cycle) or low-frequency (interannual decadal driven large-scale dynamics) may encounter different sensitivities change. We investigated this hypothesis generating groundwater level projections using deep learning models for annual, inertial (low-frequency dominated) mixed annual/low-frequency northern France from 16 CMIP6 model inputs an ensemble approach. Generated were then analysed changes variability. Generally, levels tended decrease all scenarios across the 2030-2100. The showed slightly increasing but decreasing types. As severity scenario increased, more inertial-type stations appeared be affected Focusing on confirmed observation: while significant amount less severe SSP 2-4.5 scenario, eventually slight yet statistically as increased. For almost Finally, seemed, instances, higher than historical period, without any differences between emission scenarios.

Язык: Английский

---